4. How do you design for the Cloud?
The Old Way
Share scarce system resources
The New Way
Exploit abundant system resources
.
5. Reimagining the Relational Database
What if you were inventing the database today?
You wouldn’t design it the way we did in 1970. At least not entirely.
You’d build something that can scale out, that is self-healing, and that
leverages existing AWS services.
9. Microservice Architecture
Moved the logging and storage layer into
a multi-tenant, scale-out database-
optimized storage service
Integrated with other AWS services like
Amazon EC2, Amazon VPC, Amazon
DynamoDB, Amazon SWF, and Amazon
Route 53 for control plane operations
Integrated with Amazon S3 for continuous
backup with 99.999999999% durability
Control PlaneData Plane
Amazon
DynamoDB
Amazon SWF
Amazon
Route 53
Logging + Storage
SQL
Transactions
Caching
Amazon S3
for a DB
10. Aurora Storage
Highly available by default
• 6-way replication across 3 AZs
• 4 of 6 write quorum
• Automatic fallback to 3 of 4 if an
Availability Zone (AZ) is unavailable
• 3 of 6 read quorum
SSD, scale-out, multi-tenant storage
• Seamless storage scalability
• Up to 64 TB database size
• Only pay for what you use
Log-structured storage
• Many small segments, each with their own redo logs
• Redo logs used to generate data pages on demand
• Eliminates chatter between database and storage
SQL
Transactions
AZ 1 AZ 2 AZ 3
Caching
Amazon S3
12. Do fewer IOs
Minimize network packets
Cache prior results
Offload the database engine
DO LESS WORK
Process asynchronously
Reduce latency path
Use lock-free data structures
Batch operations together
BE MORE EFFICIENT
How did we make this fast ?
DATABASES ARE ALL ABOUT I/O
NETWORK-ATTACHED STORAGE IS ALL ABOUT PACKETS/SECOND
HIGH-THROUGHPUT PROCESSING DOES NOT ALLOW CONTEXT SWITCHES
13. IO Traffic in RDS MySQL
BINLOG DATA DOUBLE-WRITELOG FRM FILES
T Y P E O F W R I T E
MYSQL WITH STANDBY
EBS mirrorEBS mirror
AZ 1 AZ 2
Amazon S3
EBS
Amazon Elastic
Block Store (EBS)
Primary
Instance
Standby
Instance
1
2
3
4
5
Issue write to EBS – EBS issues to mirror, ack when both done
Stage write to standby instance using DRBD
Issue write to EBS on standby instance
IO FLOW
Steps 1, 3, 5 are sequential and synchronous
This amplifies both latency and jitter
Many types of writes for each user operation
Have to write data blocks twice to avoid torn writes
OBSERVATIONS
780K transactions
7,388K I/Os per million txns (excludes mirroring, standby)
Average 7.4 I/Os per transaction
PERFORMANCE
30 minute SysBench writeonly workload, 100GB dataset, RDS SingleAZ, 30K PIOPS
14. IO Traffic in Aurora (Database)
AZ 1 AZ 3
Primary
Instance
Amazon S3
AZ 2
Replica
Instance
AMAZON AURORA
ASYNC
4/6 QUORUM
DISTRIBUTED
WRITES
BINLOG DATA DOUBLE-WRITELOG FRM FILES
T Y P E O F W R I T E S
30 minute SysBench writeonly workload, 100GB dataset
IO FLOW
Only write redo log records; all steps asynchronous
No data block writes (checkpoint, cache replacement)
6X more log writes, but 9X less network traffic
Tolerant of network and storage outlier latency
OBSERVATIONS
27,378K transactions 35X MORE
950K I/Os per 1M txns (6X amplification) 7.7X LESS
PERFORMANCE
Boxcar redo log records – fully ordered by LSN
Shuffle to appropriate segments – partially ordered
Boxcar to storage nodes and issue writes
15. IO Traffic in Aurora (Storage Node)
LOG RECORDS
Primary
Instance
INCOMING QUEUE
STORAGE NODE
S3 BACKUP
1
2
3
4
5
6
7
8
UPDATE
QUEUE
ACK
HOT
LOG
DATA
BLOCKS
POINT IN TIME
SNAPSHOT
GC
SCRUB
COALESCE
SORT
GROUP
PEER TO PEER GOSSIPPeer
Storage
Nodes
All steps are asynchronous
Only steps 1 and 2 are in foreground latency path
Input queue is 46X less than MySQL (unamplified, per node)
Favor latency-sensitive operations
Use disk space to buffer against spikes in activity
OBSERVATIONS
IO FLOW
① Receive record and add to in-memory queue
② Persist record and ACK
③ Organize records and identify gaps in log
④ Gossip with peers to fill in holes
⑤ Coalesce log records into new data block versions
⑥ Periodically stage log and new block versions to S3
⑦ Periodically garbage collect old versions
⑧ Periodically validate CRC codes on blocks
16. IO Traffic in Aurora (Read Replica)
PAGE CACHE
UPDATE
Aurora Master
30% Read
70% Write
Aurora Replica
100% New Reads
Shared Multi-AZ Storage
MySQL Master
30% Read
70% Write
MySQL Replica
30% New Reads
70% Write
SINGLE-THREADED
BINLOG APPLY
Data Volume Data Volume
Logical: Ship SQL statements to Replica
Write workload similar on both instances
Independent storage
Can result in data drift between Master and
Replica
Physical: Ship redo from Master to Replica
Replica shares storage. No writes performed
Cached pages have redo applied
Advance read view when all commits seen
MYSQL READ SCALING AMAZON AURORA READ SCALING
17. Asynchronous group commit acks
Read
Write
Commit
Read
Read
T1
Commit (T1)
Commit (T2)
Commit (T3)
LSN 10
LSN 12
LSN 22
LSN 50
LSN 30
LSN 34
LSN 41
LSN 47
LSN 20
LSN 49
Commit (T4)
Commit (T5)
Commit (T6)
Commit (T7)
Commit (T8)
LSN GROWTH
Durable LSN at head-node
COMMIT QUEUE
Pending commits in LSN order
TIME
GROUP
COMMIT
TRANSACTIONS
Read
Write
Commit
Read
Read
T1
Read
Write
Commit
Read
Read
Tn
TRADITIONAL APPROACH AMAZON AURORA
Maintain a buffer of log records to write out to disk
Issue write when buffer full or time out waiting for writes
First writer has latency penalty when write rate is low
Request I/O with first write, fill buffer till write picked up
Individual write durable when 4 of 6 storage nodes ACK
Advance DB Durable point up to earliest pending ACK
18. Re-entrant connections multiplexed to active threads
Kernel-space epoll() inserts into latch-free event queue
Dynamically size threads pool
Gracefully handles 5000+ concurrent client sessions on r3.8xl
Standard MySQL – one thread per connection
Doesn’t scale with connection count
MySQL EE – connections assigned to thread group
Requires careful stall threshold tuning
CLIENTCONNECTION
CLIENTCONNECTION
LATCH FREE
TASK QUEUE
epoll()
MYSQL THREAD MODEL AURORA THREAD MODEL
Adaptive thread pool
19. Improvements over the past few months
Write batch size tuning
Asynchronous send for read/write I/Os
Purge thread performance
Bulk insert performance
BATCH OPERATIONS
Failover time reductions
Malloc reduction
System call reductions
Undo slot caching patterns
Cooperative log apply
OTHER
Binlog and distributed transactions
Lock compression
Read-ahead
CUSTOMER FEEDBACK
Hot row contention
Dictionary statistics
Mini-transaction commit code path
Query cache read/write conflicts
Dictionary system mutex
LOCK CONTENTION
21. Aurora Storage Availability
Quorum system for read/write; latency tolerant
Peer to peer gossip replication to fill in holes
Continuous backup to S3 (designed for 11 9s durability)
Continuous scrubbing of data blocks
Continuous monitoring of nodes and disks for repair
10GB segments as unit of repair or hotspot rebalance to
quickly rebalance load
Quorum membership changes do not stall writes
AZ 1 AZ 2 AZ 3
Amazon S3
22. Traditional databases
Have to replay logs since the last
checkpoint
Typically 5 minutes between checkpoints
Single-threaded in MySQL; requires a
large number of disk accesses
Amazon Aurora
Underlying storage replays redo records
on demand as part of a disk read
Parallel, distributed, asynchronous
No replay for startup
Checkpointed Data Redo Log
Crash at T0 requires
a re-application of the
SQL in the redo log since
last checkpoint
T0 T0
Crash at T0 will result in redo logs being
applied to each segment on demand, in
parallel, asynchronously
Instant crash recovery
23. Survivable caches
We moved the cache out of the
database process
Cache remains warm in the event of a
database restart
Lets you resume fully loaded
operations much faster
Instant crash recovery + survivable
cache = quick and easy recovery from
DB failures
SQL
Transactions
Caching
SQL
Transactions
Caching
SQL
Transactions
Caching
Caching process is outside the DB process
and remains warm across a database restart
24. Faster, more predictable failover
App
RunningFailure Detection DNS Propagation
Recovery Recovery
DB
Failure
MYSQL
App
Running
Failure Detection DNS Propagation
Recovery
DB
Failure
AURORA WITH MARIADB DRIVER
1 5 - 2 0 s e c
3 - 2 0 s e c
25. ALTER SYSTEM CRASH [{INSTANCE | DISPATCHER | NODE}]
ALTER SYSTEM SIMULATE percent_failure DISK failure_type IN
[DISK index | NODE index] FOR INTERVAL interval
ALTER SYSTEM SIMULATE percent_failure NETWORK failure_type
[TO {ALL | read_replica | availability_zone}] FOR INTERVAL interval
Simulate failures using SQL
To cause the failure of a component at the database node:
To simulate the failure of disks:
To simulate the failure of networking:
27. Simplify Database Management
Create a database in minutes
Automated patching
Push-button scale compute
Continuous backups to Amazon S3
Automatic failure detection and failover
Amazon RDS
28. Simplify Storage Management
Read replicas are available as failover targets—no data loss
Instantly create user snapshots—no performance impact
Continuous, incremental backups to Amazon S3
Automatic storage scaling up to 64 TB—no performance or availability
impact
Automatic restriping, mirror repair, hot spot management, encryption
29. Simplify Data Security
Encryption to secure data at rest
• AES-256; hardware accelerated
• All blocks on disk and in Amazon S3 are encrypted
• Key management via AWS KMS
SSL to secure data in transit
Network isolation via Amazon VPC by default
No direct access to nodes
Supports industry standard security and data
protection certifications
Storage
SQL
Transactions
Caching
Amazon S3
Application
31. 4 client machines with 1,000 connections each
WRITE PERFORMANCE READ PERFORMANCE
Single client machine with 1,600 connections
MySQL SysBench
R3.8XL with 32 cores and 244 GB RAM
SQL Benchmark Results
33. Beyond benchmarks
If only real world applications saw benchmark performance
POSSIBLE DISTORTIONS
Real world requests contend with each other
Real world metadata rarely fits in data dictionary cache
Real world data rarely fits in buffer cache
Real world production databases need to run HA
34. Scaling Table Count
Tables
Amazon
Aurora
MySQL
I2.8XL
local SSD
MySQL
I2.8XL
RAM disk
RDS
MySQL
30K IOPS
(single AZ)
10 60,000 18,000 22,000 25,000
100 66,000 19,000 24,000 23,000
1,000 64,000 7,000 18,000 8,000
10,000 54,000 4,000 8,000 5,000
SysBench write-only workload
1,000 connections
Default settings
11x
U P TO
FA STER
35. Scaling Data Set Size
DB Size Amazon Aurora
RDS MySQL
30K IOPS (single AZ)
1GB 107,000 8,400
10GB 107,000 2,400
100GB 101,000 1,500
1TB 26,000 1,200
67x
U P TO
FA STER
SYSBENCH WRITE-ONLY
DB Size Amazon Aurora
RDS MySQL
30K IOPS (single AZ)
80GB 12,582 585
800GB 9,406 69
CLOUDHARMONY TPC-C
136x
U P TO
FA STER
36. Scaling User Connections
SysBench OLTP Workload
250 tables
Connections Amazon Aurora
RDS MySQL
30K IOPS (single AZ)
50 40,000 10,000
500 71,000 21,000
5,000 110,000 13,000
8x
U P TO
FA STER
37. Running with Read Replicas
Updates per
second Amazon Aurora
RDS MySQL
30K IOPS (single AZ)
1,000 2.62 ms 0 s
2,000 3.42 ms 1 s
5,000 3.94 ms 60 s
10,000 5.38 ms 300 s
SysBench Writeonly Workload
250 tables
500x
U P TO
LOWER LAG